Multiscale Attention Large Kernel Convolutional Neural Network for Rolling Bearing Fault Diagnosis in Noise Environment

Abstract Rolling bearings, as one of the important sources of faults in various mechanical equipment, play a significant role in engineering applications for diagnosing bearing faults in noisy environments. In light of the limited diagnostic precision and substantial computational demands associated with traditional deep learning fault diagnosis methods in noisy conditions, a novel multiscale attention large kernel convolutional neural network (SJ-CNN) model is introduced. The multiscale learning strategy uses convolution kernels of different sizes to capture features at different scales and extract features that are not affected by noise at a specific scale. By combining multiscale learning techniques with large kernel convolution operations, the model achieves a wider perception domain and can identify key fault features even when the original signal is distorted by noise. Furthermore, an adaptive attention module is designed to reassign adaptive weights to each scale to improve the attention to fault information. Additionally, two SJ-CNN variants are utilized: the first SJ-CNN (SJ-1-CNN), which excels in diagnostic accuracy and is particularly advantageous for diagnosing bearings in intricate settings; and the second SJ-CNN (SJ-2-CNN), which achieves a lightweight model and reduces significant computational costs. The effectiveness of this approach has been validated using the bearing datasets from Paderborn University (PU) and the Spectra Quest (SQ) experimental platform. Specifically, under strong noise conditions, the diagnostic accuracy of SJ-CNN has improved by 7.50% on average, which is superior to that of existing state-of-the-art methods. The code will be open in https://github.com/su-yibei/fault-diagnosis.